14 / Research / Jun 2026
AI-Generated Face Familiarity Study
A browser-based psychophysics study asking whether AI-generated faces feel familiar to people, and whether machines agree.
- AI-generated faces rated
- 60
- Familiarity scale
- 1 to 5
- Undergraduates selected for the school
- 20
Overview
The CVR-VISTA Summer School at York University's Centre for Vision Research selected 20 undergraduates for a week of vision science in the age of AI: research talks, panels, lab tours and hands-on project time spanning neuroscience, psychology and computer vision.
For our group project, my team asked a question that feels urgent as AI-generated imagery gets harder to tell apart from reality: do AI-generated faces feel familiar or memorable to us, and do machines agree?
The study
AI face generators learn from huge sets of human faces, so their output may look prototypical enough to feel familiar. To test that, we designed a browser-based psychophysics study in which participants rated 60 AI-generated faces on a 1 to 5 familiarity scale. We then compared those ratings with a neural network's familiarity predictions and with ViTMem memorability scores.
Findings
Human ratings showed a reproducible signal, and the model's predictions correlated with them. Faces people found least familiar tended to be the most memorable. We also found a possible prototypicality bias: the most familiar faces skewed male and Caucasian-coded, while women's faces appeared across a wider range of features.
